Energy Cost of Enhancing Reliability of Machine Learning Models for Edge IoT Security

Machine learning (ML) models are widely used for detecting malicious traffic in networked systems, including Internet of Things (IoT) environments. Recent work has shown that accuracy alone is insufficient, as models must also provide reliable and well-calibrated confidence estimates to avoid overconfident errors and unstable behavior in practical deployments. Improving model calibration typically relies on post-hoc techniques or additional training data, but both can increase model complexity and energy demand, which is in turn an important limitation for resource-constrained edge deployments. While the trade-off between predictive accuracy and energy consumption has been explored in prior work, the joint trade-off involving prediction reliability has received far less attention despite its critical importance for operational Intrusion Detection Systems (IDS). In this paper, we address this gap by investigating the relationship between predictive performance, reliability, and energy consumption in ML-based IDS for IoT networks. To this end, we develop an evaluation framework that applies post-hoc calibration techniques, namely, Platt scaling and isotonic regression, to ML classifiers and systematically quantifies their impact. Using real IoT traffic data, we evaluate models under varying training data sizes and measure predictive accuracy, calibration quality, and computational energy requirements. Experimental results show that post-hoc calibration improves confidence reliability considerably, with isotonic regression reducing calibration error (ECE) by over 95% and Platt scaling by 83%, while increasing inference energy by only$\mathbf{2}-\mathbf{3} {\%}$, demonstrating that reliability does not come at meaningful computational cost. Results further highlight that calibrated models achieve the same levels of predictive performance and reliability using up to$60-90 {\%}$less training data, reducing both data needs and energy demand.